Date of Award
August 2020
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Automotive Engineering
Committee Member
Yunyi Jia
Committee Member
Zoran Filipi
Committee Member
Ioannis Karamouzas
Committee Member
Bing Li
Abstract
Multi-target tracking in public traffic calls for a tracking system with automated track initiation and termination facilities in a randomly evolving driving environment. Besides, the key problem of data association needs to be handled effectively considering the limitations in the computational resources on-board an autonomous car. The challenge of the tracking problem is further evident in the use of high-resolution automotive sensors which return multiple detections per object. Furthermore, it is customary to use multiple sensors that cover different and/or over-lapping Field of View and fuse sensor detections to provide robust and reliable tracking. As a consequence, in high-resolution multi-sensor settings, the data association uncertainty, and the corresponding tracking complexity increases pointing to a systematic approach to handle and process sensor detections.
In this work, we present a multi-target tracking system that addresses target birth/initiation and death/termination processes with automatic track management features. These tracking functionalities can help facilitate perception during common events in public traffic as participants (suddenly) change lanes, navigate intersections, overtake and/or brake in emergencies, etc. Various tracking approaches including the ones based on joint integrated probability data association (JIPDA) filter, Linear Multi-target Integrated Probabilistic Data Association (LMIPDA) Filter, and their multi-detection variants are adapted to specifically include algorithms that handle track initiation and termination, clutter density estimation and track management. The utility of the filtering module is further elaborated by integrating it into a trajectory tracking problem based on model predictive control.
To cope with tracking complexity in the case of multiple high-resolution sensors, we propose a hybrid scheme that combines the approaches of data clustering at the local sensor and multiple detections tracking schemes at the fusion layer. We implement a track-to-track fusion scheme that de-correlates local (sensor) tracks to avoid double counting and apply a measurement partitioning scheme to re-purpose the LMIPDA tracking algorithm to multi-detection cases. In addition to the measurement partitioning approach, a joint extent and kinematic state estimation scheme are integrated into the LMIPDA approach to facilitate perception and tracking of an individual as well as group targets as applied to multi-lane public traffic. We formulate the tracking problem as a two hierarchical layer. This arrangement enhances the multi-target tracking performance in situations including but not limited to target initialization(birth process), target occlusion, missed detections, unresolved measurement, target maneuver, etc. Also, target groups expose complex individual target interactions to help in situation assessment which is challenging to capture otherwise.
The simulation studies are complemented by experimental studies performed on single and multiple (group) targets. Target detections are collected from a high-resolution radar at a frequency of 20Hz; whereas RTK-GPS data is made available as ground truth for one of the target vehicle's trajectory.
Recommended Citation
Hunde, Andinet Negash, "Multi Sensor Multi Target Perception and Tracking for Informed Decisions in Public Road Scenarios" (2020). All Dissertations. 2692.
https://tigerprints.clemson.edu/all_dissertations/2692